體外受精與胚胎植入（In Vitro Fertilization and Embryo Transfer， IVF-ET）為現代不孕症治療之主要手段，然即便引入著床前染色體篩選技術（PGT-A）確認胚胎染色體正常，整倍體囊胚之著床失敗率仍居高不下。現有人工智慧輔助胚胎評估研究雖已確立深度學習於影像預測之臨床潛力，然普遍面臨三項方法論困境：小樣本場景下之過擬合風險、純影像模型對患者臨床資訊之忽略，以及多模態融合中各模態之融合權重完全隱含於神經網路參數之中，無法直接量化各模態對預測結果之獨立貢獻，臨床可解釋性受到限制。 針對上述困境，本研究提出一套 CNN-SVM 混合架構之多模態預測框架，以 200 筆整倍體囊胚資料為實驗基礎，整合 ResNet-18 遷移學習微調提取 512 維深度視覺特徵，並以系統性網格搜尋自動決定年齡縮放係數之顯式加權融合機制，將患者取卵年齡（OPU Age）與影像特徵整合，最終以徑向基函數核支持向量機（RBF-SVM）進行分類決策。為系統性驗證架構設計之合理性，本研究另納入隨機森林（Random Forest）與梯度提升樹（XGBoost）作為對照分類器，構成多分類器比較實驗，全程採巢式分層五折交叉驗證確保效能評估之統計無偏性。 實驗結果顯示，於最佳條件（原始特徵＋年齡）下，SVM 之袋外累積平均準確率為 67.00%，平均召回率為 67.00%，為三種分類器中唯一準確率與召回率同步達到最佳且完全平衡者，其全量袋外累積分類報告於雙類別呈現完全對稱之判別表現（FHB＋ 與 FHB− 之 F1-Score 均為 0.67），確認本系統不具傾向性偏誤。於年齡特徵之貢獻方面，加入取卵年齡後三種分類器之準確率均呈正向提升，且各分類器於各折次所選出之最佳年齡權重均大於 1，此資料驅動結果提供了整倍體純化條件下母體取卵年齡仍具獨立預測貢獻之實證。本研究所提出之方法論框架，為後續以臨床可及樣本規模進行大規模複製驗證之研究奠定基礎。;In vitro fertilization and embryo transfer (IVF-ET) is the primary treatment modality for modern infertility management. However， even when chromosomal normality is confirmed through preimplantation genetic testing for aneuploidies (PGT-A)， implantation failure among euploid blastocysts remains prevalent. Although existing artificial intelligence-assisted embryo assessment studies have established the clinical potential of deep learning for image-based prediction， they commonly face three methodological challenges: overfitting risk in small-sample settings， the inability of image-only models to incorporate patient-level clinical information， and the inability of existing multimodal fusion frameworks to directly quantify the independent predictive contribution of each modality， as fusion weights are entirely embedded within neural network parameters and remain uninterpretable. To address these challenges， this study proposes a multimodal predictive framework based on a CNN-SVM hybrid architecture. Using 200 PGT-A-confirmed euploid blastocyst samples， the framework integrates ResNet-18 transfer learning fine-tuning to extract 512-dimensional deep visual features， and employs an explicit weighted fusion mechanism in which an age scaling coefficient is automatically determined through systematic grid search， integrating the patient′s age at oocyte pick-up (OPU Age) with image features. Classification decisions are performed by a radial basis function support vector machine (RBF-SVM). To systematically validate the rationale of the architectural design， random forest and XGBoost were additionally incorporated as comparative classifiers， forming a multi-classifier comparison experiment. Nested stratified five-fold cross-validation was adopted throughout to ensure statistically unbiased performance estimation. Experimental results demonstrate that under the optimal condition (original features with age fusion)， SVM achieved a mean out-of-fold accuracy of 67.00% and a mean recall of 67.00%， making it the only classifier among the three to attain simultaneously optimal and fully balanced accuracy and recall. Its cumulative out-of-fold classification report revealed perfectly symmetric discrimination across both classes (F1-Scores of 0.67 for both FHB+ and FHB−)， confirming the absence of directional prediction bias. Regarding the contribution of the age feature， the incorporation of OPU age improved the accuracy of all three classifiers， and the optimal age weights selected across folds were consistently greater than one， providing data-driven evidence that maternal OPU age retains an independent predictive contribution to euploid blastocyst implantation outcomes even after chromosomal confounders have been eliminated through PGT-A. The methodological framework established in this study lays the foundation for subsequent large-scale replication studies conducted at clinically accessible sample sizes.
